Electrical and Computer Engineering Department University of Waterloo Canada

Transcription

1 Predicting a Biological Response of Molecules from Their Chemical Properties Using Diverse and Optimized Ensembles of Stochastic Gradient Boosting Machine By Tarek Abdunabi and Otman Basir Electrical and Computer Engineering Department University of Waterloo Canada

2 Outline Introduction. Stochastic gradient boosting. Data and computational resources. Building and tuning the ensembles. Performance evaluation. Fusing ensembles decisions. Conclusion and future work. Supplementary material (R code).

3 Introduction The development of a new drug largely depends on trial and error. It typically involves synthesizing thousands of compounds that finally becomes a drug. As a result, this process is extremely expensive and slow. Therefore, the ability to accurately predict the biological activity of molecules, and understand the rationale behind those predictions are of great value.

4 Stochastic gradient boosting Gradient Boosting Machines (GBMs) are powerful ensemble learning techniques. In boosting methods, new models are added sequentially to the ensemble. At each iteration, a new base-leaner model is trained with respect to the error of the ensemble learned in the previous iterations. Despite their high accuracy, GBMs suffer from major drawbacks such as high memory consumption.

5 Data and computational resources The data was obtained from the Kaggle.com s competition: Predicting a Biological Response held between March16, 2012 and June 15, The objective of the competition was to build a predictive model to optimally relate molecular information to an actual biological response. The first column contains experimental data describing an actual biological response (Active/Inactive). The remaining columns represent molecular descriptors (D1 through D1776) e. g. Size, shape, etc.

6 Building and tuning the ensembles We build optimized GBMs ensembles using several feature selection/reduction techniques. The performance of each ensemble is compared to the performance of the optimized ensemble built using all predictors (1776). Then, we use two fusion techniques to obtain better prediction accuracy.

7 Building and tuning the ensemble using all predictors

8 Building and tuning the ensemble using PCA (first 255 with 95% var.)

9 Building and tuning the ensemble using PCA (Kaiser-Guttman rule)

10 Building and tuning the ensemble using predictors area under ROC

11 Building and tuning the ensemble using the Relief algorithm

12 Building and tuning the ensemble using GBM s predictor importance

13 Performance evaluation In-sample performance

14 Performance evaluation Out-of-sample performance

15 Performance evaluation Computations time The time required to build and tune our ensembles using Amazon cloud computing (32 cores) was: Ensemble 1 (all predictors) ~ 43.2 hours Ensemble 2 (PCA 255 comp.) ~ 2.9 hours Ensemble 3 (PCA Kaiser) ~ 3.58 minutes Ensemble 4 (ROC area) ~ minutes Ensemble 5 (Relief score) ~ minutes Ensemble 6 (GBM var. Impor.) ~ 1.1 hours

16 Fusing the ensembles decisions Two techniques were used, Majority Vote and an optimized feedforward neural network. For unbiased results, we used stratified random sampling to split the test dataset to training and testing datasets. The prediction accuracy of ensembles 1 to 6 are 79.53%, 78.04%, 72.7%, 79.28%, 79.23%, and 79.53% respectively. The prediction accuracy using the majority vote is 79.53% ( better than ensembles 2 to 5, and equals to ensembles 1 and 6)

17 Fusing the ensembles decisions Predictions of ensembles 2-6 on the training dataset are used to build and tune a feedforward neural network. Prediction accuracy on the test data is 80.12%

18 Conclusion and future work The ability to accurately predict the biological activity of molecules, and understand the rationale behind those predictions are of great value in drug discovery. Using different feature selection/reduction techniques, diverse and optimized Tree-based Gradient Boosting Machines were built for a real, high-dimensional dataset.

19 Conclusion and future work By using these techniques, the computations time for building an optimized GBM using all predictors is significantly reduced at a slight drop in prediction accuracy. A better prediction accuracy is obtained by fusing ensembles decisions using a majority vote and an optimized feedforward neural network. For future work, more feature selection, models, and fusion techniques will be investigated.

20 Supplementary material (R code) Complete R code and data analysis steps Or Tarek Abdunabi at: tabdunab@uwaterloo.ca